Deep Learning–Based Surface Crack Identification Of Reinforce Concrete Structures

Cracks are one of the common damages of civil structures.The main method of traditional crack detection is artificial visual inspection,which is labor-intensive and subjective in practical applications.In recent years,the development of computer vision and deep learning theory has provided a more automated means for crack detection and identification.However,when image processing techniques detect cracks in an image,features need to be extracted.Thus,the use of image processing techniques is also limited because images taken on real concrete surfaces are subject to noise caused by illumination,blur,and the like.This paper studies the identification method of surface cracks in concrete structures based on deep learning.The main research contents of this paper are as follows:Based on the generalization theory of deep learning and the stochastic gradient descent algorithm,an actual concrete surface crack detection method based on convolutional neural network is proposed.The convolutional neural network(CNN)framework and each network layer are designed.Node connection,etc.,to achieve automatic extraction of automatic crack image features.The design of CNN was trained and tested using a consumer-grade camera to capture surface cracks in concrete structures.To train CNN,an image database is first created,a large number of images are taken from the real concrete surface,and cut into small images using the window sliding method,sorted and marked.An architecture was built to detect cracks using CNN training data sets.Several databases containing different numbers of sub-images were used to compare the accuracy and time of the training.Most databases contain 86,000 sub-images with an image size of 96 x 96 pixels,which achieves 98% accuracy with the proposed structure.UAV was used to image the surface crack of the actual concrete frame structure,and the feasibility of using the image acquired by the UAV for crack identification was discussed.The validity of the proposed CNN-based crack identification method was further verified.